Author puts his script before an AI jury: the verdict cost 850 thousand tokens

Language models flatter. Ask whether your idea is good and you will almost always hear "great." On October 29, the author of the Telegram channel Technokafe tested an idea differently.

He installed the free Claude Court skill for Claude Code (Anthropic's coding agent) and put the script of his own video on trial. A skill is a ready-made set of instructions that extends what an agent can do. The result: the script was acquitted by two votes out of three, and the review took 8 agents and 849 660 tokens.

A token here is the unit in which a model counts the volume of text. The number of tokens determines both the price and how much of a subscription limit gets used.

How a court of neural networks works

The skill stages a trial. Each role is played by a separate agent, meaning a separate language-model session.

• The prosecutor looks for what is wrong with the idea. • The defender argues that it is viable. • Jurors get roles (for example, statistician) and each votes on their own. They cannot see each other, so they do not adjust to anyone else's opinion. • The judge writes the final file, VERDICT.md.

The point of the file is not praise. The judge lists the weak points and states the condition under which the idea can be launched. The skill's code is on GitHub: alexyc9381/court-skill.

An ordinary "so, how is it?" gets one point of view, and that view tends to agree with you. Here there are sides whose job is to argue. There is also independent voting: jurors do not read each other, so they do not copy a neighbor's mistakes.

What the court found and what the judge demanded

The script was acquitted 2:1. The statistician voted against. His objection: the plan did not contain a single original number. The judge accepted the argument and set a condition: first measure the cost, then shoot the video.

The author measured it and got these figures:

• A trial with three jurors: 8 agents, 849 660 tokens. • A full trial with twelve jurors: 17 agents, about 1.8 million tokens. This is a calculation, not an actual run. • The trimmed-down panel costs half as much as the full one, not a quarter. The prosecution, the defense and the judge work regardless of the number of jurors, so that part of the cost stays constant.

These numbers show one more thing, though this is our own estimate. 849 660 tokens for 8 agents and 1.8 million for 17 work out to roughly 100 thousand tokens per agent in both cases. Each additional juror adds about that much. If you want a cheap trial, cut jurors. Keep the judge and the two opposing sides.

When a trial pays off

The author is upfront about where this applies. For a minor edit, such a trial is unnecessary. For an idea that will take a week of work, 850 thousand tokens are cheaper than a week wasted. The logic is simple: checking an idea costs tokens, while a planning mistake costs days. The same logic explains why the statistician proved more useful than the others. He demanded that vague wording be replaced with a measurable number.

What the video itself cost

The author also gives the cost of the rest of the pipeline:

• Claude Code on Opus 5.5 built the script, the edit and the style. Including the style test, that comes to about $13 in API-equivalent terms (at pay-per-token rates) and 1.3% of the weekly subscription limit. • The voice is the author's own, a clone in ElevenLabs. 960 characters cost $0.11. • Sixteen figures were drawn by GPT Image through Codex on a ChatGPT subscription. • The animation was made with HyperFrames and GSAP.

The $13 does not include the trial; the author measured its cost separately, in tokens. Still, the order of magnitude is clear. The whole video, with voiceover, graphics and a courtroom-style review of the idea, was assembled from subscriptions and API calls that cost pennies.

Practical takeaways

Engineers and product teams can draw several lessons from this story.

• One question to one model yields a compliment. A structured argument with independent voting yields a list of weak points. • Before launching a multi-agent review, measure its cost. The author did exactly that, at the judge's demand. • If the budget is limited, reduce the number of jurors rather than removing the judge and the two sides. • It makes sense to use such a court for decisions where a mistake is expensive, such as plans for a week of work. For minor edits it is overkill.

The method's limitation is obvious. All participants in the court are still language models and may share the same blind spots. Even with independent voting, the result remains the opinion of models, not a reality check. Even so, a verdict with specific conditions is more useful than "great idea!", and it costs less than a million tokens.